A Modular Approach to Training Cascades of Boosted Ensembles

Abstract

Building on the ideas of Viola-Jones [1] we present a framework for training cascades of boosted ensembles (CoBE) which introduces further modularity and tractability to the training process. It addresses the challenges faced by CoBE frameworks such as protracted runtimes, slow layer convergences and classifier optimization. The framework possesses the ability to bootstrap positive samples and may in turn be extended into the domain of incremental learning. This paper aims to address our framework's susceptibility to overfitting with possible solutions. Experiments are conducted on face detectors using the bootstrapping of large positive datasets and their accuracy, with respect to overfitting, is examined.

Adaptive cascade of boosted ensembles for face detection in concept drift

“…The comparisons were conducted between our algorithm and the same underlying classifier that was trained on stationary data without concept learning capabilities. The classifier in question was first trained offline using the [29] method with Haar-like features. Five thousand positive faces from a combination of FERET 1 and Yale Face Database B [31] data sets were used, against a α α α α Fig.…”

“…The idea of a dual-cascaded structure was taken further in Susnjak et al [29], in order to implement positive sample bootstrapping. This enabled the utilization of potentially massive positive data sets, without the learning algorithm being exposed to all samples explicitly, a task that would otherwise be computationally too expensive.…”